mHC框架
Search documents
西部证券晨会纪要-20260106
Western Securities· 2026-01-06 01:37
Group 1: Computer Industry Insights - DeepSeek has released the mHC framework, which indicates the evolution direction for next-generation infrastructure, enhancing large model training stability, speed, and cost-effectiveness [3][4] - The mHC architecture introduces manifold constraints, utilizing the Sinkhorn-Knopp algorithm to project residual mapping matrices onto Birkhoff polytope, demonstrating superior performance and scalability in large-scale training [3][4] - The mHC framework is expected to inspire new research paths and improve the balance between plasticity and stability in model training, potentially leading to innovative methods in architecture design [4] Group 2: AI Chip Design Paradigm - The mHC framework proposes a new paradigm for AI chip design by addressing the mismatch between computing power and bandwidth, advocating for a shift from "computing power-first" to "efficiency-first" in chip design [5] - mHC's optimization techniques, such as kernel fusion and selective recomputation, significantly reduce bandwidth requirements by consolidating multiple memory accesses into a single access [5] - The introduction of specialized projection operator acceleration units in chip design could disrupt the current dominance of general-purpose computing units in AI chips, promoting a heterogeneous architecture [5] Group 3: Non-ferrous Metals Industry Insights - The non-ferrous metals sector remains promising, with China's manufacturing PMI for December 2025 exceeding expectations, indicating an overall economic recovery [9] - The U.S. initial jobless claims for the week ending December 27, 2025, were lower than expected, suggesting a strengthening labor market [10] - Geopolitical tensions have escalated following U.S. military actions in Venezuela, which may impact the non-ferrous metals market due to increased security risks [11] Group 4: North Exchange Market Insights - The North Exchange market is expected to see a spring rally, supported by high-quality expansion and liquidity improvements, despite short-term profit-taking pressures [14][16] - Recent policy initiatives, including the digital transformation plan for the automotive industry, are anticipated to provide opportunities for high-end manufacturing enterprises in the North Exchange [16] - The North Exchange's valuation remains attractive compared to the Sci-Tech Innovation Board, suggesting a favorable investment environment for institutional investors [16]
DeepSeek发布mHC框架,或为下一代基础架构指明演进方向
Western Securities· 2026-01-05 08:20
Investment Rating - The industry investment rating is "Overweight" [2][9] Core Insights - DeepSeek has introduced the mHC framework, which enhances the stability, speed, and cost-effectiveness of large model training [2] - The mHC architecture is a significant extension of the Hyper-Connections (HC) paradigm, paving the way for future research and innovations in model training [3] - The mHC framework combines manifold constraints with engineering optimizations, potentially revolutionizing AI chip design by addressing the mismatch between computing power and bandwidth [4] Summary by Sections Industry Overview - The mHC framework is expected to provide substantial performance improvements in large-scale model training, demonstrating stability and efficiency [2][3] Technological Advancements - The introduction of manifold constraints in the mHC framework allows for various explorations tailored to specific learning objectives, which may lead to novel methods balancing flexibility and stability [3] - The mHC framework's approach to software adapting to hardware bottlenecks challenges traditional chip design paradigms, promoting a shift towards efficiency [4] Investment Opportunities - The report suggests focusing on domestic AI chip companies such as Cambrian, Haiguang Information, Moer Thread, Muxi Co., and Birun Technology as potential investment opportunities [5]